Cognitive assistant for aiding expert decision

1. A method comprising:

evaluating electronic health records of a first patient to identify state changes, wherein each respective state change corresponds to a point in time when one or more medical attributes of the first patient changed;

identifying, in the electronic health records, a first plurality of decisions made by a first healthcare provider during treatment of the first patient, wherein each respective decision of the plurality of decisions corresponds to a respective state change, and wherein at least a first decision of the plurality of decisions corresponds to the first healthcare provider taking no action and continuing with a treatment plan in response to a state change;

determining, for each respective decision of the first plurality of decisions, one or more corresponding medical attributes of the first patient that were present at a time when the respective decision was made;

generating a trajectory for the first patient, wherein the trajectory reflects the plurality of decisions and, for each respective decision of the first plurality of decisions, the corresponding medical attributes;

training a cognitive assistant, using a generative adversarial imitation learning (GAIL) model, based on the trajectory, comprising:

training a generator neural network to generate medical decisions based on the trajectory, comprising:

weighting each respective decision of the first plurality of decisions based at least in part on recency of each respective decision, wherein more recent decisions are provided with increased weight; and

weighting each respective decision of the first plurality of decisions based on a respective subject matter expert (SME) that made the respective decision, wherein a higher weight is provided for a first set of decisions of the first plurality of decisions, as compared to a weight provided for a second set of decisions of the first plurality of decisions, based on determining that the first set of decisions were made by a particular SME and the second set of decisions were made by a different SME, wherein both the first set of decisions and the second set of decisions are used to train the generator neural network; and

training a discriminator neural network to differentiate between generated medical decisions and the first plurality of decisions;

receiving one or more medical attributes of a second patient; and

generating a first medical decision by processing the one or more medical attributes of the second patient using the cognitive assistant.

2. The method of claim 1 , wherein monitoring the plurality of decisions made by the first healthcare provider comprises analyzing electronic health records (EHR) of the first patient to identify each of the first plurality of decisions.

3. The method of claim 1 , wherein generating the first medical decision is performed upon determining that a cohort of the second patient matches a cohort of the first patient.

4. The method of claim 3 , the method further comprising:

receiving one or more medical attributes of a third patient;

upon determining that a cohort of the third patient does not match the cohort of the first patient, identifying a second cognitive assistant that was trained for the cohort of the third patient; and

generating a second medical decision by processing the one or more medical attributes of the third patient using the second cognitive assistant.

5. The method of claim 1 , wherein generating the first medical decision comprises suggesting a first lab test for the second patient.

6. The method of claim 5 , the method further comprising:

receiving a first result for the first lab test for the second patient; and

generating a second medical decision by processing the first result using the cognitive assistant, wherein generating the second medical decision comprises suggesting a second lab test for the second patient.

7. The method of claim 1 , wherein training a cognitive assistant comprises, for each respective decision of the first plurality of decisions:

generating a training decision by providing the corresponding one or more medical attributes of the first patient that were present at the time when the respective decision was made to the cognitive assistant;

identifying the respective decision; and

refining the cognitive assistant based on comparing the training decision with the respective decision.

8. The method of claim 1 , wherein the cognitive assistant is trained based on medical decisions by the first healthcare provider with respect to a plurality of patients.

9. The method of claim 1 , wherein the cognitive assistant is further trained based on medical decisions by a plurality of healthcare providers with respect to a plurality of patients.

10. A computer program product comprising:

a computer-readable storage medium having computer-readable program code embodied therewith, the computer-readable program code executable by one or more computer processors to perform an operation comprising:

evaluating electronic health records of a first patient to identify state changes, wherein each respective state change corresponds to a point in time when one or more medical attributes of the first patient changed;

identifying, in the electronic health records, a first plurality of decisions made by a first healthcare provider during treatment of the first patient, wherein each respective decision of the plurality of decisions corresponds to a respective state change, and wherein at least a first decision of the plurality of decisions corresponds to the first healthcare provider taking no action and continuing with a treatment plan in response to a state change;

determining, for each respective decision of the first plurality of decisions, one or more corresponding medical attributes of the first patient that were present at a time when the respective decision was made;

generating a trajectory for the first patient, wherein the trajectory reflects the plurality of decisions and, for each respective decision of the first plurality of decisions, the corresponding medical attributes;

training a cognitive assistant, using a generative adversarial imitation learning (GAIL) model, based on the trajectory, comprising:

training a generator neural network to generate medical decisions based on the trajectory, comprising:

weighting each respective decision of the first plurality of decisions based on a respective subject matter expert (SME) that made the respective decision, wherein a higher weight is provided for a first set of decisions of the first plurality of decisions, as compared to a weight provided for a second set of decisions of the first plurality of decisions, based on determining that the first set of decisions were made by a particular SME and the second set of decisions were made by a different SME, wherein both the first set of decisions and the second set of decisions are used to train the generator neural network; and

weighting each respective decision of the first plurality of decisions based on a respective subject matter expert (SME) that made the respective decision, wherein a higher weight is provided for a first set of decisions of the first plurality of decisions, as compared to a weight provided for a second set of decisions of the first plurality of decisions, based on determining that the first set of decisions were made by a particular SME and the second set of decisions were made by a different SME, wherein both the first set of decisions and the second set of decisions are used to train the generator neural network; and

training a discriminator neural network to differentiate between generated medical decisions and the first plurality of decisions;

receiving one or more medical attributes of a second patient; and

generating a first medical decision by processing the one or more medical attributes of the second patient using the cognitive assistant.

11. The computer program product of claim 10 , wherein generating the first medical decision is performed upon determining that a cohort of the second patient matches a cohort of the first patient.

12. The computer program product of claim 11 , the operation further comprising:

receiving one or more medical attributes of a third patient;

upon determining that a cohort of the third patient does not match the cohort of the first patient, identifying a second cognitive assistant that was trained for the cohort of the third patient; and

generating a second medical decision by processing the one or more medical attributes of the third patient using the second cognitive assistant.

13. The computer program product of claim 10 , wherein generating the first medical decision comprises suggesting a first lab test for the second patient.

14. The computer program product of claim 13 , the operation further comprising:

receiving a first result for the first lab test for the second patient; and

generating a second medical decision by processing the first result using the cognitive assistant, wherein generating the second medical decision comprises suggesting a second lab test for the second patient.

15. The computer program product of claim 10 , wherein training a cognitive assistant comprises, for each respective decision of the first plurality of decisions:

generating a training decision by providing the corresponding one or more medical attributes of the first patient that were present at the time when the respective decision was made to the cognitive assistant;

identifying the respective decision; and

refining the cognitive assistant based on comparing the training decision with the respective decision.

16. A system comprising:

one or more computer processors; and

a memory containing a program which when executed by the one or more computer processors performs an operation, the operation comprising:

evaluating electronic health records of a first patient to identify state changes, wherein each respective state change corresponds to a point in time when one or more medical attributes of the first patient changed;

identifying, in the electronic health records, a first plurality of decisions made by a first healthcare provider during treatment of the first patient, wherein each respective decision of the plurality of decisions corresponds to a respective state change, and wherein at least a first decision of the plurality of decisions corresponds to the first healthcare provider taking no action and continuing with a treatment plan in response to a state change;

determining, for each respective decision of the first plurality of decisions, one or more corresponding medical attributes of the first patient that were present at a time when the respective decision was made;

generating a trajectory for the first patient, wherein the trajectory reflects the plurality of decisions and, for each respective decision of the first plurality of decisions, the corresponding medical attributes;

training a cognitive assistant, using a generative adversarial imitation learning (GAIL) model, based on the trajectory, comprising:

training a generator neural network to generate medical decisions based on the trajectory, comprising:

weighting each respective decision of the first plurality of decisions based on a respective subject matter expert (SME) that made the respective decision, wherein a higher weight is provided for a first set of decisions of the first plurality of decisions, as compared to a weight provided for a second set of decisions of the first plurality of decisions, based on determining that the first set of decisions were made by a particular SME and the second set of decisions were made by a different SME, wherein both the first set of decisions and the second set of decisions are used to train the generator neural network; and

weighting each respective decision of the first plurality of decisions based on a respective subject matter expert (SME) that made the respective decision, wherein a higher weight is provided for a first set of decisions of the first plurality of decisions, as compared to a weight provided for a second set of decisions of the first plurality of decisions, based on determining that the first set of decisions were made by a particular SME and the second set of decisions were made by a different SME, wherein both the first set of decisions and the second set of decisions are used to train the generator neural network; and

training a discriminator neural network to differentiate between generated medical decisions and the first plurality of decisions;

receiving one or more medical attributes of a second patient; and

generating a first medical decision by processing the one or more medical attributes of the second patient using the cognitive assistant.

17. The system of claim 16 , wherein generating the first medical decision is performed upon determining that a cohort of the second patient matches a cohort of the first patient, the operation further comprising:

receiving one or more medical attributes of a third patient;

upon determining that a cohort of the third patient does not match the cohort of the first patient, identifying a second cognitive assistant that was trained for the cohort of the third patient; and

generating a second medical decision by processing the one or more medical attributes of the third patient using the second cognitive assistant.

18. The system of claim 16 , wherein generating the first medical decision comprises suggesting a first lab test for the second patient.

19. The system of claim 18 , the operation further comprising:

receiving a first result for the first lab test for the second patient; and

generating a second medical decision by processing the first result using the cognitive assistant, wherein generating the second medical decision comprises suggesting a second lab test for the second patient.

20. The system of claim 16 , wherein training a cognitive assistant comprises, for each respective decision of the first plurality of decisions:

generating a training decision by providing the corresponding one or more medical attributes of the first patient that were present at the time when the respective decision was made to the cognitive assistant;

identifying the respective decision; and

refining the cognitive assistant based on comparing the training decision with the respective decision.